Drop-in agent team orchestration for Claude Code. Zero dependencies. Zero scripts.

A CLAUDE.md + .claude/ config that turns Claude Code's native Agent Teams into a structured orchestration system with defined roles, model routing, and automated quality gates.

Before (this repo, old version): ~2,500 lines of TypeScript + Bash scripts reimplementing what Claude Code already provides natively.

Now: 3 config files, 320 lines total. Everything runs through Claude Code's built-in Agent Teams feature.

# Clone this repo
git clone https://github.com/Devgapperk/Claude-Code-Guide.git

# Copy the orchestration files into your project
cp Claude-Code-Guide/CLAUDE.md /path/to/your-project/
cp -r Claude-Code-Guide/.claude/ /path/to/your-project/

cd /path/to/your-project
claude

Tell Claude what to build:

Create an agent team to build user authentication.
Spawn an architect (opus), two engineers (sonnet), a validator (haiku), and a scribe (haiku).
Use delegate mode.

Claude reads CLAUDE.md, spawns teammates with the right prompts and models, creates the task list, and orchestrates.

your-project/
├── CLAUDE.md                          # Orchestration blueprint (loaded by all teammates)
└── .claude/
    ├── settings.json                  # Enables agent teams + permissions + hooks
    └── hooks/
        ├── stop-loop-guard.sh         # Detects Ralph Loops (agent spinning on same error)
        ├── task-completed.sh          # Blocks task completion if tests/types fail
        ├── teammate-idle.sh           # Prevents idle if uncommitted work exists
        ├── post-edit-verify.sh        # Type-checks after every file edit (catches slop)
        └── block-destructive.sh       # Blocks rm -rf, force push, DROP TABLE

1. CLAUDE.md is auto-loaded by every teammate on spawn. It defines roles, model routing, anti-slop protocol, loop prevention rules, context window management, and conventions.

2. settings.json enables agent teams, pre-approves safe commands (test, lint, build, git), blocks dangerous ones (.env reads, rm -rf), and wires up 5 hooks across the full lifecycle.

3. Hooks enforce quality at every stage:

   Hook	Event	What It Does
   stop-loop-guard.sh	Stop	Scans transcript for repeated errors or retry language. If an agent hit the same error 3+ times or retried 4+ times, forces it to stop and report instead of burning more tokens.
   task-completed.sh	TaskCompleted	Runs npm test + tsc --noEmit. Blocks completion if either fails. Teammate gets the error output and must fix it.
   teammate-idle.sh	TeammateIdle	Checks for uncommitted changes. Keeps teammate working until they commit.
   post-edit-verify.sh	PostToolUse (Write/Edit)	Type-checks after every .ts/.tsx file change. Catches hallucinated imports and wrong signatures on the first edit, not after 10 files of broken code.
   block-destructive.sh	PreToolUse (Bash)	Blocks rm -rf, git push --force, and DROP TABLE.

Only spawn what the task needs. A bug fix might only need 2 engineers + validator.

Full feature:

Create an agent team to build [feature]. Spawn architect (opus), frontend
engineer (sonnet), backend engineer (sonnet), validator (haiku), scribe (haiku).
Require plan approval for the architect. Use delegate mode.

Parallel code review:

Create an agent team to review PR #142. Three reviewers: security, performance,
test coverage. Have them report independently.

Bug investigation:

Users report [symptom]. Create a team with 4 teammates investigating competing
hypotheses. Have them challenge each other's theories.

• Claude Code CLI
• jq (used by hook scripts)
• Git

No npm install. No build step. No runtime dependencies.

Built by DevGap. Orchestrated with Kleiber. Powered by Claude.